-
Notifications
You must be signed in to change notification settings - Fork 7
/
Chapter_5_questions
130 lines (101 loc) · 8.81 KB
/
Chapter_5_questions
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
Shared Memory Computer
Kysymys osion loppuun
The fundamental feature of a shared-memory computer is that all the CPU cores are connected to the same part of memory. Which of the following statements about
shared-memory computer holds true?
• The shared-memory architecture is only utilized in supercomputers (FALSE)
• Memory capacity, memory access speed and race conditions are key challenges of the shared-memory approach (TRUE)
• Even if the connection between the CPU and the memory eventually becomes a bottleneck adding more CPU cores still brings significant advantages in performance (FALSE)
• Parallel programming with shared memory is extremely difficult and thus usually avoided. (FALSE)
Distributed Memory
Kysymys osion loppuun
All modern supercomputers use the same basic approach in building large systems: many separate computers with own memories are connected with a fast network.
Which of the statements concerning distributed memory computers are true?
• In distributed memory model, each separate computer (or node) operates independently but can directly access the memory in other nodes. (FALSE)
• It is much easier to build networks connecting great quantities of computers instead of having numerous CPU cores in a single shared-memory computer. (TRUE)
• Programming distributed-memory computer can be challenging (TRUE)
• Seemingly infinite capacity and computing power, controlled usage of the memory bus are benefits of the distributed memory model (TRUE)
Memory Hierarchy
In computers, there is a hierarchy of different levels of memory. Below you’ll find few statements about memory hierarchy, and your task is to select all that apply.
• A memory cache is basically a large amount of memory far from the CPU-core. In general, a memory cache is a synonym for the main memory. (FALSE)
• Cache works extremely well in reading data but there are some limitations/challenges (cache coherency) to be concerned in writing data. (TRUE)
• Memory types closest to the processor have the highest performance (transfer speed in MB/s) but at the same time the lowest capacity (information storage) (TRUE)
• Accessing main memory requires communication via the interconnect (FALSE)
Parallelization with CPU cores
CPU performance has improved tremendously over the years, while the memory performance has increased at a much slower pace. To improve the performance, amount of
reads and writes to the memory must be optimised. Optimisation is done in program code. Select the methods that can be used to improve performance with better
memory use.
• = Structure the code to enable compiler instruction level parallelism.
• = Vectorizing parts of code where possible.
• ~ Reducing amount of cycles by giving less instructions.
• ~ Overclocking the supercomputer's random access memory (RAM).
Graphics processing units
GPU (Graphics processing unit) is a central component in modern high-performance computing. Select all the factual statements from the options below
• GPUs are extremely useful in machine learning applications such as training neural networks. (TRUE)
• CPU is a more complex and more suitable for general purpose usage than a GPU (TRUE)
• In a near future supercomputers will most probably have only GPUs because GPUs are more energy efficient than CPUs and can operate efficiently without CPUs. (FALSE)
• Extracting good performance out of a GPU can be a true challenge for a programmer as the low-level hardware details might need to be considered (TRUE)
• Until now there haven’t been GPUs available for consumer markets, but the recent development in supercomputing has led to a situation where there will soon be
GPUs also for personal computers. (FALSE).
Interconnect
• Kirjoitusvirheitä tekstissä, korjattava!!
Many scientific computing problems are tightly coupled (lots of communication needed), thus a high-speed interconnect is essential for solving them. Which of the
following statements are true?
• Fast interconnect network between nodes is equally important for normal desktop computers as it is for supercomputers (FALSE)
• The bandwidth of interconnect should be as small as possible (FALSE)
• MAHTI supercomputer operated by CSC utilizes a fully connected network topology, in which there is a direct connection between all pairs of nodes. (FALSE)
• When choosing an interconnect topology, there are lots of parameters that need to be considered (number of connections, cost, performance, distance between
nodes, etc.) – (TRUE)
How Fast Can a Supercomputer Be?
Thera are several factors affecting the performance of a supercomputer. Which of the following statements are true?
• In theory interconnect could dramatically limit the performance of a supercomputer (TRUE)
• Even in best cases, it is possible to reach roughly up to 50 % of the theoretical peak performance of a supercomputer (FALSE)
• The performance is affected by algorithmic choices, programming techniques and type of a computational problem (TRUE)
Chapter 5 – lopputentti
1. Construct the memory hierarchy pyramid. Place the memory/storage types so that the fastest is on top and the lowest at the bottom.
• Registers
• L1
• L2
• L3
• Main memory (RAM)
• Remote memory
• File system disks
2. Select all that apply. A distributed memory model tries to…
• = solve memory and memory access limitations of a shared memory architecture
• = improve scalability of capacity and computing power
• ~ make supercomputer software building easier
• ~ offer faster access to memory for CPUs
3. Select all the correct statements that are related to interconnect
• = To gain better performance, different nodes must be located relatively near to each other.
• = All things taken into consideration, in the most optimized synthetic scenario performance is up to 80 % of the theoretical peak performance of a supercomputer.
• ~ Mahti's interconnect network topology is awesomely and called Dragon's flight based on the TV serie Game of Thrones.
• ~ All of Mahti's nodes are connected to each other through a fully connected network
• ~ Interconnect is the second fastest part of the chain feeding data to CPU's, right after L1 cache.
4. There are several types of memory available in supercomputers. Connect the terms and their definitions.
• Registers = All the arithmetic operations on data are performed here. Physically located on a CPU core.
• L1, L2, L3 = intermediate caches. When the CPU core needs to fetch data, it looks first from here.
• Main memory = locates within a node. All instructions and data of active programs are stored here.
• Remote memory = Accessing requires communication via the interconnect.
• Disks = Store data which doesn't fit into main memory. Slow access.
5. GPU (Graphics processing unit) is a central component in modern high-performance computing. Select all the correct statements from the options below
• GPUs are extremely useful in machine learning applications such as training neural networks. (TRUE)
• GPU is a more complex and more suitable for general purpose usage than a CPU (FALSE)
• GPUs need CPUs on their side (TRUE)
• Extracting good performance out of a GPU can be a true challenge for a programmer as the low-level hardware details might need to be considered (TRUE)
• GPUs can easily be used as accelerators in all scientific problems (FALSE)
6. The fundamental feature of a shared-memory computer is that all the CPU cores are connected to the same piece of memory. Which of the following statements about
shared-memory computer is true?
• The shared-memory architecture is only utilized in supercomputers (FALSE)
• Memory capacity, memory access speed and race conditions are key challenges of the shared-memory approach (TRUE)
• Even if the connection between the CPU and the memory eventually becomes a bottleneck adding more CPU cores still brings significant advantages in
performance (FALSE)
• Single-memory approach is preferred in supercomputing as parallel programming with shared memory is extremely difficult (FALSE)
7. CPU performance has improved tremendously over the years, while the memory performance has increased at a much slower pace. To improve the performance, amount
of reads and writes to the memory must be optimised. Optimisation is done in program code. Select the methods that can be used to improve performance
with better memory use.
• = Structure the code to enable compiler instruction level parallelism.
• = Vectorizing parts of code where possible.
• ~ Reducing amount of cycles by giving less instructions.
• ~ Overclocking the supercomputer's random access memory (RAM).
8. In supercomputers, there are typically four to six GPUs and one to two CPUs per node. In favourable cases, GPUs can typically speed up the software by a factor of four to eight
(or even more) when compared to situation where only CPUs are utilized, true or false?
TRUE